Re: Bug: fio traps into kernel without exiting because futex has a deadloop

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2009-06-11 at 16:33 +0800, Zhang, Yanmin wrote:
> On Thu, 2009-06-11 at 08:18 +0200, Peter Zijlstra wrote:
> > On Thu, 2009-06-11 at 07:55 +0200, Peter Zijlstra wrote:
> > > On Thu, 2009-06-11 at 11:08 +0800, Zhang, Yanmin wrote:
> > > > I investigate a fio hang issue. When I run fio multi-process
> > > > testing on many disks, fio traps into kernel and doesn't exit
> > > > (mostly hit once after runing sub test cases for hundreds of times).
> > > > 
> > > > Oprofile data shows kernel consumes time with some futex functions.
> > > > Command kill couldn't kill the process and machine reboot also hangs.
> > > > 
> > > > Eventually, I locate the root cause as a bug of futex. Kernel enters
> > > > a deadloop between 'retry' and 'goto retry' in function futex_wake_op.
> > > > By unknown reason (might be an issue of fio or glibc), parameter uaddr2
> > > > points to an area which is READONLY. So futex_atomic_op_inuser returns
> > > > -EFAULT when trying to changing the data at uaddr2, but later get_user
> > > > still succeeds becasue the area is READONLY. Then go back to retry.
> > > > 
> > > > I create a simple test case to trigger it, which just shmat an READONLY
> > > > area for address uaddr2.
> > > > 
> > > > It could be used as a DOS attack.
> > 
> > /me has morning juice and notices he sent the wrong commit...
> > 
> > commit 64d1304a64477629cb16b75491a77bafe6f86963
> > Author: Thomas Gleixner <tglx@linutronix.de>
> > Date:   Mon May 18 21:20:10 2009 +0200
> 2.6.30 includes the new commit. I did a quick testing with my simple
> test case and it traps into kernel without exiting.
> 
> The reason is I use flag FUTEX_PRIVATE_FLAG. So the fshared part in function
> get_futex_key should be deleted. That might hurt performance.

FWIW, using a private futex on a shm section is wrong in and of itself.

tglx: should we create CONFIG_DEBUG_FUTEX and do a vma lookup to
validate that private futexes are indeed in private anonymous memory?

But you would be able to trigger the same using an PROT_READ anonymous
mmap().

It appears access_ok() isn't as strict as we'd like it to be:

/*
...
 * Note that, depending on architecture, this function probably just
 * checks that the pointer is in the user space range - after calling
 * this function, memory access functions may still return -EFAULT.
 */
#define access_ok(type, addr, size) (likely(__range_not_ok(addr, size) == 0))

Thomas is working on a fix for this.